A vague request leaves room for a wrong answer. Explore how language models work, add useful context, and learn why a fluent response still needs checking.
Your activity / Right here
Bring a notebook. Begin below.
The activity is included on this page. Work on paper or in your own document; keep your first attempt and the evidence you checked.
Warm-up: Dr. Q's Question
Why did Mike bring back the wrong thing?
My first answer: ________________________________________________
Mike's Mix-Up: The Study Guide for the Wrong Test
Made-Up Example.
The night before his biology test, Mike typed three words into a chatbot: "make study guide." The AI produced a long, neat guide to the whole cell: organelles, mitosis, and DNA. Mike studied it for twenty minutes before he remembered that his test was on photosynthesis.
"This AI is useless," he texted Carla.
Carla wrote back: "Or did you ask it for the thing from the other thing?"
Here is what Mike did:
1. He typed a request with no subject, topic, date, level, or format.
2. He accepted the first answer without comparing it with his class notes.
3. He studied from it for twenty minutes.
4. He blamed the tool.
What went wrong? Talk it through with a partner, then write your answers.
• Which layers of context were missing? Name at least four.
• Using Rob's equation, where did the error come from?
• What should Mike have checked before he started studying?
• Rewrite Mike's request as a Level 3 request.
Think First: The Thing From the Other Thing
No devices. A partner, about ten small objects, and paper. About 15 minutes.
1. Put about ten small objects on a table: pencils, pens, an eraser, a notebook, a paper clip.
2. Partner A secretly picks one object and writes it down.
3. Partner A gives the vaguest request possible: "Get me the thing from the other thing."
4. Partner B picks something. If it is wrong, Partner A adds one layer of context, such as a color, a place, or a size.
5. Keep going until Partner B gets it. Count the layers and the wrong guesses.
6. Switch roles and play two more rounds.
On paper, chart layers of context along the bottom and wrong guesses up the side. Write one sentence about what happens to the wrong guesses as the layers go up.
Read the video transcript
Read along, pause, or return to an idea. This transcript follows the selected lesson video.
Welcome back.
I'm Carla, an AI avatar.
Today we find out how AI actually works.
It starts with a pencil.
Hey Mike, can you please get me the thing from the other thing?
Mike brings back a stapler.
Why did Mike bring back the wrong thing?
Dr. Q tries again.
"Grab the pencil from that thing."
Wrong pencil.
"Grab the pencil from the desk."
Closer.
Then: "Grab the red pencil from the bucket on top of the
stack of books that is on my desk."
Mike gets it in two seconds.
The robot gets it too.
So does the AI app on Dr. Q's phone.
A clear request gets the right answer.
Notice what changed between the first request and the
last: the words.
Every word Dr. Q added ruled out a wrong guess.
Rob, the author of our textbook,
calls this the mission: reduce the error.
When you prompt AI, you are building a prediction.
The output is the AI's prediction,
based on the context you gave it.
The gap between what you need and what you get is the error.
Here is why.
Chatbots are built on large language models.
They read text in small pieces called tokens.
For English, one hundred tokens are about seventy-five words.
In training, the model plays a giant fill-in-the-blank game.
Predict the next word.
Check.
Adjust.
Repeat, across huge amounts of text.
Then people rank the model's answers,
and it is trained toward the answers people prefer.
That is called reinforcement learning from human feedback.
When you type a prompt, the model predicts its answer one
token at a time, using your words as context.
Vague words, big error.
Layers of context, smaller error.
Some tools also search the web or read files you upload.
That adds more context, and it can add new mistakes too,
so the audit step still applies.
Even with great context, AI can make things up.
OpenAI researchers wrote that "language models hallucinate
because standard training and evaluation procedures reward
guessing over acknowledging uncertainty."
Their example: asked for one researcher's birthday,
a chatbot gave three different dates,
and all three were wrong.
A birthday does not follow a pattern,
so the model guessed.
So your job has two parts.
Give context to shrink the error.
Then audit to catch what is left.
This week you'll play the thing from the other thing game
with a partner.
No phones needed.
Before you use an AI tool
Follow your course or institution's rules. Use an instructor-approved tool that permits your age, and meet any required consent conditions. Keep names, contact details, school records, and other private information out of prompts. Use fictional examples when you can.
Learning without an AI account
Follow the demonstration, pause before each answer, and do the reasoning on paper. Discuss what you would ask, how you would check it, and which decisions would remain yours.
About the characters and voices
This course uses AI-created characters and synthetic voices. Carla is an AI avatar. They help tell the story; they are not real students sharing testimonials.
Designed for high-school and community-college learners. Choose examples from your classes, work, and everyday life. Keep your completed work in your course portfolio.
